v0.10.2 — Scoped Chat, Method Guide & Persistence Fixes

Follow-up release to v0.10.1: scoped chat, the first #117 deliverables, and a batch of community-reported fixes.

💬 Scoped Chat

• he talk --source/--tag (and ka.chat) narrow the retrieved context to attributed or tagged documents, with capability detection for non-tracked KA types. (#161, #160)

📖 Choosing a Method (steps 1 & 2 of #117)

• New "baseline vs structured" guide in the docs. (#162)
• Side-by-side example: chunk_rag vs graph_rag on the same corpus with cost/latency comparison. (#163)

🏗️ Internal

• GraphIndexMixin extracts the shared merge/index/search surface of graph & hypergraph (−168 lines). (#145)

🐛 Fixes

• Temporal/spatial observation context now survives dump/load (was silently lost). (#167)
• MCP load failures return as tool results instead of crashing the stdio session. (#169)
• he config init no longer writes a broken same-provider embedder for LLM-only providers (Anthropic/DeepSeek/Gemini). (#165)
• Cog_RAG.chat forwards source_ids/tags to search. (#171)
• he list --lang zh keeps method templates. (#159)
• Document templates get output-shape validation. (#173)
• CLI banner lists jsonld/cypher. (#155)

📚 Docs

• Resolved leftover conflict markers in MCP docs; unsmashed JSON-LD/Cypher sections; type-aware index-layout notes; Gemini provider guide. (#149, #151, #153, #157)

Full changelog: v0.10.1...v0.10.2